METHODOLOGY Data Sources

Data used for this research will be gathered from the Bureau of Economic Analysis (BEA) and the Annual Survey of Manufacturers (ASM). From BEA, we use an input-output account table, which shows how much a manufacturing industry’s output was used as an input of another industry. Specifically, the table shows how manufacturing industries provide inputs and use outputs from each other to produce gross domestic product. By using this table, we will be able to determine supplying industries to a focal manufacturer and industries that are supplied from a focal manufacturer. As a result, we are able to create a map of the supply base and customer base for any given industry. In addition, we will use ASM data which provides knowledge of investments in IT infrastructure, including investments in computers and

peripheral equipment and software and communication equipment. ASM also provides data on costs, inventories, total cost of materials, total value of shipments, value added, and production wages.

The I/O table is released every 5 years with the newest one being from 2007 (BEA 2013). In this study we will create a complexity index using the I/O table from the year of 2007. Due to different NAISC codes between ASM and BEA, the BEA format is used and thus, aggregates some industries’ observations in the ASM data to match with BEA data.

100

Table 3.1. Variables and Measurements

Type Variable/ Proxy of Measures Source

Independent HHIxSupCount/ Supply base complexity (#  )1  ∑ * +, ∑ , +,- . /01 2 3456789#93:7 BEA Independent HHIxCustCount/ Customer base complexity #(#  )1  ∑ * +, ∑ , +,- . /01 2 #3;56789#93:7 BEA

Dependent Gross margin/ profitability

(Value added- production wages)/ total value of shipment

ASM

Dependent Inventory turns/ Inventory performance

RMI= raw mat inv/ total shipment FGI= fin. goods inv/ total shipment Total= Σ(RMI, WIP, FGI)/ total shipment

ASM

Moderating IT expenses/ IT infrastructure

IT= Σ (expensed computer hardware and equipment, purchases of software and data processing, other purchased computer services)

ASM

Control Adv. Intensity, Industry size, durable/ non durable ASM, BEA

Variables

Independent Variables

Supply base complexity is calculated using the formula below:

(#  <1  = )_{∑ >}>/ / / 2 . /01 ? 3456789#93:7

This formula captures the two characteristics of SCB complexity. First, it captures the degree of dispersionbetween an industry and its supplier industries as measured by the reversed Herfindahl Index (HI). The original HI shows if an industry is concentrated or widely dispersed.

101

When reversed, the index will be scaled from 0 to 1 with 0 indicating a concentrated supply base and 1 indicating a dispersed and competitive supply base. In the first component, i denotes the focal manufacturer, j denotes the supplier industries, xij denotes the total value of supplies from industry j to industry i. Second, it captures the number of suppliers that an industry interacts with. To be consistent with the first component, we divided by the total available industries in the list, thus creating a ratio. Furthermore, the index will also be easily interpreted, with a low index indicating low complexity while a high index shows high complexity. Note that by combining the two components, the scale remains from 0 to 1. Using the input-output table, we are able to identify all of the supplier industries and their shares of input for a focal manufacturer (xij). In this study, we restrict the supplier industries to only those that supply raw materials. This study argues that raw materials that focal industries use as input make up the value of outputs produced by manufacturing industries. SBC is log-transformed before estimation.

Customer base complexity is calculated using the formula below:

#(#  <1  = )_{∑ @}@/ / / 2 . /01 ? #3;756789#93:7

The formula for customer base complexity is calculated the same was as that for supply base complexity. The only difference is that instead of looking across columns in an input-output table, we look across rows to identify all of the industries that a focal manufacturer supplies, with the addition of end consumer expense, government expense, government investment, and net exports that make up GDP (yij). For calculating the customer base, we include all industries and other GDP measures as those are all customer industries to a focal manufacturer. CBC is log- transformed before estimation.

102

IT is calculated as the sum of the capital expenditures of computer hardware and equipment, purchases of software, and data processing and other purchased computer services (Yao and Zhu 2012). The data for calculating IT is provided in the Annual Survey of

Manufacturers (ASM).

Dependent Variables

There are two types of performance measures that we use here. First is Return on Sales (ROS) as a measure of financial performance. ROS is calculated as (value added-SG&A)/ total value of shipment. According to a U.S. Census, value added measures the difference between sales and total costs of materials. Removing labor costs and selling, general, and administrative expenses from the value added yields to operating income. This calculation is consistent with those in the literature (Cachon et al. 2007, Yao and Zhu 2012). Second, we calculated inventory performance using raw materials inventory (RMI), and finished goods inventory (FGI). RMI is calculated as raw materials inventory divided by total value of shipments. Finally, FGI is calculated as finished goods inventory divided by divided by total value of shipments. All these measures are consistent with the literature (Saldanha et al. 2013). All dependent variables are log-transformed, except return on sales.

Control Variables

There are three control variables used in this study. The first is Advertising Intensity (AdsIn) calculated as advertising expenses over value of product shipments. The level of

advertising intensity should correlate with the inventory level and sales and thus, an appropriate control variable for this study. The second variable is durable goods industry which is measured

103

using a dummy variable. The annual survey of manufacturers groups industries into two

categories: durable goods and non-durable goods industry. Literature suggests that the impacts of durable goods vs. nondurable goods industries on margin and inventory level are significantly different (Yao and Zhu 2012). In particular, nondurable goods industries tend to have a higher margin than the durable goods industries. Finally, this study also controls industry size which is calculated as log natural of value of product shipments.

Model Specification

The dependent variables in the model are Return on Sales (ROS), raw material inventory turnover (RMI_turn), and finished goods inventory turnover (FGI_turn). The main explanatory variables are supply base complexity (SBC), customer base complexity (CBC), and interaction of SBC and CBC (SBCxCBC). Control variables are industry size, advertising intensity, and

durable goods industry. The first set of models (equations 1-3) estimates the impact of supply base complexity and customer base complexity on performance. The second set of models (equations 4-6) estimates the interaction effect between SBC and CBC and its impact on performance. The moderating effect of IT is estimated using the second model (equations 4-6) and a median split approach. Note that our dependent variables may be correlated to one another; that is, finished goods are product of raw material goods, and they both affect ROS particularly with regards to inventory cost. Accordingly, the errors of the three equation models in each set may be correlated to one another.

The main effect model has the following form:

ROS = β1+ β2SBC+ β3CBC + β4-6Control_Vars + ε (1)

104

FGI_turn = β21+ β22SBC+ β23CBC + β24-26Control_Vars + ε (3)

The interaction and moderation effect model has the following form:

ROS = β1+ β2SBC+ β3CBC + β4SBCxCBC + β5-7Control_Vars + ε (4)

RMI_turn = β11+ β12SBC+ β13CBC+ β14SBCxCBC+ β15-17Control_Vars + ε (5)

FGI_turn = β21+ β22SBC+ β23CBC+ β24SBCxCBC+ β25-27Control_Vars + ε (6)

Data

Table 3.2 presents industry characteristics and distributions of industry by industry sector (three-digit NAICS). As seen in the data, machinery, transportation equipment, and food sectors have the highest number of observations, representing one-third of the total sample. In contrast, apparel and leather sectors have only one observation each in the data. Average of the industry sector sales are about $25 billion with a wide range across industry sectors. The petroleum and coal sector has the highest average sales with $147.81 billion while leather and electrical equipment sectors have the lowest average sales with $5 billion and $7.14 billion, respectively.

In terms of supply base complexity, industries in transportation equipment, furniture, and machinery sectors have the highest while textile products and petroleum and coal sectors have the lowest complexity. In terms of customer base complexity, industries in plastic and rubber, machinery, and primary metal sectors are the most complex while industries grouped in wood, textile products, and petroleum and coal sectors are the least complex. In general, the machinery sector has a relatively more complex supply and customer base compared to other industry sectors. In contrast, industries in textile products and petroleum and coal sectors are less complex in terms of both supply and customer base compared to other sectors. In terms of IT intensity,

105

printing, miscellaneous, and leather sectors have the highest, while textile mills, food, and petroleum and coal sectors have the lowest.

Table 3.2. Sample Characteristics

NAICS Industry Description

No. of Obs. Avg. Industry Sales ($B) Average Supply Base Complexity Average Cust. Base Complexity Average IT Intensity Avg. RMI Turn Avg. FGI Turn Avg. ROS 311 Food 24 23.49 0.168 0.089 6.79E-04 28.17 21.53 0.219

312 Beverage and Tobacco 5 24.46 0.162 0.083 7.18E-04 24.74 35.04 0.404

313 Textile Mills 3 11.68 0.160 0.064 7.09E-04 20.15 16.10 0.099

314 Textile Product Mills 3 8.79 0.135 0.053 1.32E-03 14.62 12.34 0.069

315 Apparel 1 16.88 0.161 0.089 1.59E-03 18.34 8.40 -0.235

316 Leather 1 5.00 0.173 0.085 2.96E-03 10.84 6.98 -0.002

321 Wood 4 24.25 0.229 0.054 1.26E-03 15.74 16.07 0.011

322 Paper 8 21.38 0.189 0.062 1.27E-03 14.89 17.32 0.159

323 Printing 2 48.81 0.221 0.063 4.24E-03 22.84 25.36 0.073

324 Petroleum and Coal 4 147.81 0.118 0.033 4.51E-04 30.63 19.70 0.172

325 Chemical 19 34.96 0.195 0.285 1.37E-03 18.60 11.86 0.226

326 Plastic and Rubber 10 20.07 0.192 0.391 1.23E-03 16.84 14.44 0.104 327 Nonmetallic Mineral 12 10.20 0.224 0.128 9.15E-04 15.73 14.76 0.168 331 Primary Metals 10 22.60 0.167 0.322 7.68E-04 27.03 24.19 0.116 332 Fabricated Metal 20 15.95 0.223 0.269 1.58E-03 12.84 16.07 0.082

333 Machinery 30 10.74 0.256 0.368 1.86E-03 11.10 13.97 0.055

334 Computer and Electronics 20 17.93 0.239 0.170 2.87E-03 10.97 22.17 0.106 335 Electrical Equipments 17 7.14 0.206 0.297 1.18E-03 16.54 19.32 0.113 336 Transportation Equipment 25 28.94 0.272 0.171 9.14E-04 26.14 72.14 0.081

337 Furniture 8 9.86 0.258 0.157 1.81E-03 13.49 20.90 0.052

339 Miscellaneous 11 12.32 0.235 0.200 3.19E-03 11.935 8.145 0.104

Total 237

In terms of performance, petroleum and coal, food, and primary metal are among sectors that have a high RMI turnover while machinery, computer and electronics, and leather are among sectors with a low RMI turnover. In terms of FGI turnover, the transportation equipment sector has the highest with 72.14 turns per year, which is more than double the average of the FGI

106

turnover in the manufacturing industry in the U.S. In contrast, industries in apparel,

miscellaneous, and leather sectors have, on average, the lowest FGI turnover. In terms of ROS, beverage and tobacco, chemical, and food sectors have the highest ROS while leather and apparel sectors have, on average, not only the lowest, but also a negative ROS.

The average RMI turnover and FGI turnover for all industries in the sample is 17.92 and 23.15, respectively. The average ROS for this sample is 11.8% with a median ROS of 9.8%. The mean of SBC is 0.22 while the mean of CBC is 0.21. Other statistics and the correlation matrix of variables used in this study are as indicated in table 3.3.

As discussed, a median split approach was performed to distinguish the IT intensity level of industry in the sample. In particular, industries with a value of IT intensity lower than the median value are grouped in a Low IT subsample while industries with value of IT intensity higher than the median value are grouped in a High IT subsample. The median split yielded equal sample sizes with 118 observations in each subsample. Statistics of variables in each subsample are provided in table 3.4.

To avoid a multicollinearity issue as well as to avoid losing observations during the log transform, the interaction term between SBC and CBC are shifted up by the mean value of the respected variable (Modi & Mishra, 2011). In particular, 0.22 and 0.21 were added to every value of SBC and CBC, respectively. This action, which has a similar purpose to mean centering, shifts the scale over to mitigate any potential multicollinearity issue but retains the units for the regression purpose. Furthermore, adding the mean to every value, instead of subtracting the mean from every value in mean centering, provides all positive values for SBC and CBC variables for the log transformation purpose. Variance inflation factors (VIF) test for

107

multicollinearity were then run for all variables. The test shows that VIF scores for all variables are low to moderate, with the highest score being 4.88, lower than the commonly maximum acceptable level of 10 (Lunt, 2003), suggesting multicollinearity is not an issue in this study.

Table 3.3. Descriptive Statistics and Correlation Matrix Used in the Regressions

Variable Mean Std.Dev. lnRM_Turn lnFG_Turn ROS lnsbc lncbc ln_sbcxcbc AdvInt

lnRM_Turn 2.71 0.55 1 lnFG_Turn 2.76 0.72 0.53* 1 ROS 0.12 0.13 0.06 0.001 1 lnsbc -1.60 0.40 -0.21* 0.02 -0.17* 1 lncbc -2.34 1.52 -0.23* -0.12 -0.07 0.11 1 ln_sbcxcbc -1.83 0.52 -0.19* -0.06 -0.12 0.39* 0.84* 1 AdvInt (x 10-3) 4.50 6.58 -0.23* -0.36* -0.20* 0.14* -0.08 -0.02 1 lnSize 16.35 0.91 0.34* 0.28* 0.16* 0.14* -0.07 -0.03 -0.23*

* Significant correlations, with p-value < 0.05

Table 3.4. Descriptive Statistics of Median Split Subsamples

High IT Intensity Industries Low IT Intensity Industries

Variable Mean Std. Dev. Variable Mean Std. Dev.

lnRM_Turn 2.46 0.38 lnRM_Turn 2.96 0.58 lnFG_Turn 2.59 0.57 lnFG_Turn 2.94 0.80 ROS 0.08 0.13 ROS 0.15 0.13 lnsbc -1.50 0.31 lnsbc -1.69 0.46 lncbc -2.10 1.35 lncbc -2.58 1.65 ln_sbcxcbc -1.72 0.54 ln_sbcxcbc -1.93 0.49 AdvInt (x 10-3) 6.81 8.22 AdvInt (x 10-3) 2.23 3.01 lnSize 16.17 0.86 lnSize 16.51 0.92

108

RESULTS

Estimation Models and Regression Results

In this study, there are three separate regression models to test for industry performance with respect to raw material inventory turnover, finished goods inventory turnover, and return on sales. Although the three models seem to be independent of each other, the error terms of the models are expected to be correlated. Therefore, seemingly unrelated regression (SUR) is performed to estimate the regression equations. The Breusch-Pagan test of independence confirms that residuals are indeed correlated with Chi-squared of 64.592, justifying the use of Seemingly Unrelated Regression estimation technique.

To control for Type II error, a statistical power analysis is conducted by following Cohen (1992). Having enough statistical power should provide confidence to reject H0, when H0 is

false. Cohen (1992) suggests that a minimum power of 0.80 is required to minimize the risk of making a Type II error. Statistical power can be determined for given α, sample size, and effect size. As Cohen suggests, effect sizes (ES) for multiple regression analysis can be small (ES < 0.02), medium (ES < 0.15), or large (ES < 0.35) (Cohen, 1992). Given α = 0.05, Power = 0.80, 6 independent variables, and sample size of 237 for the full sample, this study is able to detect small ES. This study can also still detect small to medium ES for the IT intensity median split sample with 118 observations (Cohen, 1992).

Regression results for all models and all explanatory variables are presented in Table 3.5- Table 3.8. Overall; all models have significant Wald Chi-squared statistics. For models with a full sample, the lowest Wald Chi-squared statistic is 41.59 and highly significant (p < 0.001). For models with an IT intensity median split sample, the lowest Wald Chi-squared statistic is 14.66 (p < 0.05).

109

Table 3.5. Main Effect Estimation Result

lnRM_Turn lnFG_Turn ROS

Variables Coeff. P-value Coeff. P-value Coeff. P-value

lnsbc -0.237 0.009 -0.075 0.522 -0.012 0.601 lncbc -0.067 0.002 -0.078 0.004 -0.002 0.738 ln_sbcxcbc AdvInt -11.608 0.021 -37.038 0.000 -2.867 0.026 lnSize 0.179 0.000 0.199 0.000 0.008 0.398 Durable -0.115 0.135 0.372 0.000 -0.084 0.000 _cons -0.621 0.380 -0.868 0.345 0.029 0.874 N 237 237 237 Chi-Squared 74.57 71.64 41.59 R-sq 0.239 0.232 0.150

Table 3.6. Interaction Effect Estimation Result

lnRM_Turn lnFG_Turn ROS

Variables Coeff. P-value Coeff. P-value Coeff. P-value

lnsbc -0.372 0.000 -0.145 0.277 -0.009 0.724 lncbc -0.166 0.000 -0.130 0.018 0.000 0.985 ln_sbcxcbc 0.358 0.006 0.186 0.280 -0.007 0.831 AdvInt -11.399 0.021 -36.930 0.000 -2.872 0.025 lnSize 0.179 0.000 0.199 0.000 0.008 0.398 Durable -0.135 0.077 0.362 0.000 -0.084 0.000 _cons -0.404 0.564 -0.756 0.564 0.024 0.894 N 237 237 237 Chi-Squared 84.46 73.16 41.64 R-sq 0.262 0.236 0.150

For the main effect model, the coefficient for SBC is negative and significant (β= -0.237, p < 0.01), indicating that supply base complexity reduces raw material inventory turnover. The coefficient for CBC is also negative and significant (β = -0.067, p < 0.01), indicating that CBC also reduces raw material inventory turnover. The regression result on the FGI turn for CBC is

110

also consistent with that on the RMI turn. CBC tends to decrease finished goods inventory turnover (β = -0.078, p < 0.01). Note that the negative effect of the complexity from supply base is larger than that from the customer base. The Wald test is significant and thus, confirms the significant different effect size between SBC and CBC. In general, hypothesis 1 and hypothesis 2 are supported for the RMI turnover model, partially supported for the FGI turnover model, and not supported for the ROS model.

Figure 3.3. Interaction between SBC and CBC using Full Sample

A results summary for the interaction effect model is shown in Table 3.6. Here, the interaction effect is introduced while controlling for main effects. The coefficient for the interaction term of SBC and CBC is positive and significant (β = 0.358, p < 0.01). Because the sign of the interaction coefficient is opposite to that of SBC and CBC, the result indicates that the

1 2 3 4 5 6 7 Low SBC High SBC ln _ R M I_ T u rn o v e r